ScaFE: Data-Efficient Scar Classification with LLM-Generated Clinical Feature Programs
ScaFE turns LLM medical knowledge into local, auditable scar-feature code instead of sending patient photos to a hosted VLM.
The paper says those generated programs measure visually assessable scar attributes, while raw images and patient-level outputs stay local. A Random Forest then classifies the structured features as keloid or hypertrophic scar. In leave-one-site-out testing on 600 photos from three hospitals, ScaFE reached 81.0% site-macro balanced accuracy, 10.0 points above BiomedCLIP. With 10% of the development data, it still reported 72.0% balanced accuracy and an 11.8-point lead. ArXiv · AI/CL/LG's note
The paper says those generated programs measure visually assessable scar attributes, while raw images and patient-level outputs stay local. A Random Forest then classifies the structured features as keloid or hypertrophic scar. In leave-one-site-out testing on 600 photos from three hospitals, ScaFE reached 81.0% site-macro balanced accuracy, 10.0 points above BiomedCLIP. With 10% of the development data, it still reported 72.0% balanced accuracy and an 11.8-point lead. ArXiv · AI/CL/LG's note
score 4